Spatio-temporal mapping of Madagascar's Malaria Indicator Survey results to assess Plasmodium falciparum endemicity trends between 2011 and 2016
←
→
Page content transcription
If your browser does not render page correctly, please read the page content below
Kang et al. BMC Medicine (2018) 16:71
https://doi.org/10.1186/s12916-018-1060-4
RESEARCH ARTICLE Open Access
Spatio-temporal mapping of Madagascar’s
Malaria Indicator Survey results to assess
Plasmodium falciparum endemicity trends
between 2011 and 2016
Su Yun Kang1, Katherine E. Battle1, Harry S. Gibson1, Arsène Ratsimbasoa2,3, Milijaona Randrianarivelojosia4,5,
Stéphanie Ramboarina2,3,6, Peter A. Zimmerman6, Daniel J. Weiss1, Ewan Cameron1, Peter W. Gething1
and Rosalind E. Howes1,6*
Abstract
Background: Reliable measures of disease burden over time are necessary to evaluate the impact of interventions and
assess sub-national trends in the distribution of infection. Three Malaria Indicator Surveys (MISs) have been conducted
in Madagascar since 2011. They provide a valuable resource to assess changes in burden that is complementary to the
country’s routine case reporting system.
Methods: A Bayesian geostatistical spatio-temporal model was developed in an integrated nested Laplace
approximation framework to map the prevalence of Plasmodium falciparum malaria infection among children
from 6 to 59 months in age across Madagascar for 2011, 2013 and 2016 based on the MIS datasets. The model
was informed by a suite of environmental and socio-demographic covariates known to influence infection
prevalence. Spatio-temporal trends were quantified across the country.
Results: Despite a relatively small decrease between 2013 and 2016, the prevalence of malaria infection has increased
substantially in all areas of Madagascar since 2011. In 2011, almost half (42.3%) of the country’s population lived in areas
of very low malaria risk (20%) increased from only 2.2% in 2011 to 9.2% in
2016. A comparison of the model-based estimates with the raw MIS results indicates there was an underestimation of
the situation in 2016, since the raw figures likely associated with survey timings were delayed until after the peak
transmission season.
Conclusions: Malaria remains an important health problem in Madagascar. The monthly and annual prevalence maps
developed here provide a way to evaluate the magnitude of change over time, taking into account variability in survey
input data. These methods can contribute to monitoring sub-national trends of malaria prevalence in Madagascar as
the country aims for geographically progressive elimination.
Keywords: Madagascar, Plasmodium falciparum, Geostatistical model, Map, Malaria Indicator Surveys
* Correspondence: rosalind.howes@bdi.ox.ac.uk
1
Malaria Atlas Project, Oxford Big Data Institute, Nuffield Department of
Medicine, University of Oxford, Oxford, UK
6
Center for Global Health and Diseases, Case Western Reserve University,
Cleveland, OH, USA
Full list of author information is available at the end of the article
© The Author(s). 2018 Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0
International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and
reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to
the Creative Commons license, and indicate if changes were made. The Creative Commons Public Domain Dedication waiver
(http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated.
Kang et al. BMC Medicine (2018) 16:71 Page 2 of 15 Background tendencies in the burden of malaria in Madagascar [4], Malaria remains an important public health problem in important uncertainties associated with this stream of data Madagascar despite concerted efforts over decades to con- present difficulties with robustly quantifying spatio- trol transmission. At times, these interventions, which temporal trends. Madagascar has benefited from three have been primarily focussed on targeted vector control, Malaria Indicator Surveys (MIS 2011, 2013 and 2016), have successfully reduced transmission to very low levels which provide standardised and rigorous measures of mal- [1–3]. The island’s malaria epidemiology is highly varied, aria endemicity at high spatial resolution from representa- reflecting a diverse ecological environment, with differing tive population samples from across the island [9–11]. seasonal trends across the country [4], which can be A recent overview of Madagascar’s routine malaria further affected by unpredictable cyclonic activity that im- case data from 2010 to 2015 described the trends in pedes malaria control efforts. Epidemics have punctuated reported case data across eight ecozones, which are pro- the history of malaria in Madagascar [5]. They are closely grammatically relevant clusters of spatially contiguous associated with the spread of rice production [6] and re- districts that share similar malaria epidemiological main an important component of the disease epidemi- characteristics [4] (Fig. 1). This sub-national perspective ology [7]. Domestic political support and international offered insight into important variation in malaria bur- investment following the turn of the century led to im- den across the island’s different environmental zones. pressive reductions in transmission up to 2008, but were The overview identified an increase in numbers of followed by a period of stagnation and subsequent reversal confirmed malaria cases. This is explained in part by of progress in the aftermath of political instability in 2009 increased access to rapid diagnostic tests, but also by a and funding disbursement delays [4]. growing malaria burden, which is indicated by a signifi- The National Malaria Control Programme (NMCP) of cant rate of increase in the diagnostic positivity rate Madagascar has recently entered a new phase, marked by between 2010 (33%) and 2015 (50%) (p < 0.0001). The the inauguration of the 2018–2022 National Strategic Plan routine reported case data, therefore, indicate there was for Malaria, which includes ambitious targets for progress an increasing malaria burden from 2010 to 2015. towards elimination during this period [8]. Central to opti- The quality of data reporting through a country’s rou- mising the intervention policy strategy for the coming 5 tine health information system, however, can vary, poten- years is a thorough understanding of the current epi- tially impacting the reliability of the nationally aggregated demiological situation across the country and of the im- health statistics as malariometric indicators, especially in a pact of investments in recent years. While routinely resource-limited setting such as Madagascar, which has a reported clinical case data can provide insight into general fragile health and communications infrastructure [12]. At Fig. 1 Malaria Indicator Survey screening sites for (a) 2011, (b) 2013 and (c) 2016. The coloured regions represent the country’s eight malaria ecozones [4]
Kang et al. BMC Medicine (2018) 16:71 Page 3 of 15
every step of the data reporting chain, cases are lost from parasitology teams. Survey locations are plotted by year in
the system [13], for example from patients not interacting Fig. 1 and by month in Additional file 1: Figure S1. The
with public health facilities during malaria episodes [14], maps reveal a west to east longitudinal gradient in the
rapid diagnostic test stock-outs resulting in unconfirmed cluster sampling in both 2013 and 2016, with the high en-
infections, or accessibility difficulties for health workers demicity east coast sampled later than western and central
submitting monthly activity reports to the centralised populations. This sampling bias, with the east coast sam-
database etc. [4]. Reported numbers can, therefore, be pled later in the transmission season, reinforced the need
strongly influenced by factors external to the true under- to account for this in the model structure. Full details of
lying epidemiology. the MIS protocols are available in the original published
MIS datasets provide a complementary perspective on reports [9–11].
recent trends in malaria endemicity, free from the intrin- Environmental and socio-demographic variables known
sic limitations of reported case data. Their strength lies to interact with and influence P. falciparum prevalence
in their standardised methodology applied repeatedly [15] were assembled as 30 arcsecond spatial grids (Table 1).
over multiple years across nationally representative sites. Among the suite of covariates, eight were temporally static
This present study uses the MIS data to map the preva- covariates. The socio-demographic variables included
lence of infection by Plasmodium falciparum in those stable night-time lights in 2010 [16], which indicated the
under 5 years old (6 to 59 months in age: PfPR6–59mo) to presence of cities or towns [17], and accessibility to cities,
obtain a deeper insight into epidemiological trends than which quantified travel time to cities larger than 50,000
that available from the published MIS overview analyses, people [18]. Static environmental variables included eleva-
which reported aggregated national prevalence rates of tion as measured from the Shuttle Radar Topography
6.2% for 2011 [9], 9.1% for 2013 [10] and 7.0% for 2016 Mission (SRTM) [19], slope, which was calculated from
[11]. This study aims to generate spatially continuous that elevation in ArcGIS 10.5.1 [20], and a measure of
prevalence maps that account for disparities in sample distance to water, which indicates the Euclidean distance
collection dates between the different MIS years. This to permanent and semi-permanent water based on the
variation in the time window is not otherwise accounted presence of lakes, wetlands, rivers and streams, and ac-
for in the published MIS analyses [9–11]. The maps pro- counts for slope and precipitation [21, 22]. Also included
vide a benchmark of transmission intensity, free from were indices of aridity [23], potential evapotranspiration
the nuances of the datasets of routinely reported case [23] and topographic wetness (which was calculated from
data. These can support strategic resource allocation the SRTM elevation surface).
and allow progress towards the current strategic plan Temporally dynamic covariates are those for which data
goals to be monitored. were available at monthly intervals, or, for population
density, annual intervals [24]. Precipitation data were
Methods obtained from the Climate Hazards Group Infrared
Study data Precipitation with Station data (CHIRPS) [25]. A P. falcip-
Parasite prevalence data for this study came from MIS arum-specific covariate developed by the Malaria Atlas
activities conducted across Madagascar in 2011, 2013 Project that measures the suitability of air temperature for
and 2016. Alongside many other indicators, an MIS malaria transmission [26, 27] was included after it was
includes measurements of the prevalence of P. falcip- extended to the full study period. The remaining dynamic
arum infection among children 6 to 59 months old variables were obtained from Moderate Resolution Im-
(PfPR6–59mo), based on a standardised protocol with a aging Spectroradiometer (MODIS) satellite data [28]. Note
stratified sampling strategy across the island that is that these data were first gap-filled [29] to remove missing
designed to capture a nationally representative assess- values, such as due to cloud cover, before being aggregated
ment of prevalence. Data are freely available from the to monthly measurements. Three products of temperature
Demographic and Health Surveys (DHS) online reposi- data were derived from land surface temperature (LST),
tory. Across the three MIS years, a total of 898 spatially including (i) daytime LST, (ii) night-time LST and (iii) the
unique geo-positioned datapoints were available, which diurnal difference of daytime to night-time LST [30].
reported prevalence of infection among 19,986 children. MODIS reflectance data were used for measurements of
The initial survey in 2011 included 266 sites (29.6% vegetation and moisture: enhanced vegetation index (EVI)
overall; 6827 individuals), screened from March to May. [31], tasselled cap brightness (TCB) and tasselled cap
In 2013, 274 sites (30.5%; 6232 individuals) were screened wetness (TCW) [32].
between April and June. Finally, in 2016, 358 sites (39.9%; The values of the covariate data at the MIS screening lo-
6927 individuals) were screened from April to July. The cations were extracted for all points. Temporally dynamic
diagnostic outcomes used here were from microscopy, covariates were matched to the same month as the survey,
read by the Institut Pasteur de Madagascar and NMCP as well as 1, 2 and 3 months prior to the survey (lags).
Kang et al. BMC Medicine (2018) 16:71 Page 4 of 15
Table 1 List of environmental and socio-demographic covariates
Covariate Description Dynamic Round 1 Round 2 Source
selection selection
Accessibility Distance to cities with population >50,000 Static Yes Yes Nelson [18]
AI Aridity index Static Yes Yes World Clim [23]
DistToWater GIS-derived surface that measures distance Static Yes Yes MAP (from WWF surfaces
to permanent and semi-permanent water [21, 22])
based on presence of lakes, wetlands, rivers
and streams, and accounting for slope and
precipitation
Elevation Elevation as measured by the Shuttle Radar Static Yes Yes SRTM derivative [19]
Topography Mission (SRTM)
PET Potential evapotranspiration Static Yes Yes Trabucco & Zomer [23]
Slope GIS-derived surface calculated from SRTM Static Yes Yes MAP (from SRTM [19])
elevation surface
Stable_Lights_2010 Index that measures the presence of lights Static Yes Yes NOAA [16]
from towns, cities and other sites with
persistent lighting
TWI Topographic wetness index Static Yes Yes MAP (from SRTM [19])
Population size Estimated population per 1 km × 1 km pixel Annual Yes Linard et al. [24]
EVI Enhanced vegetation index Monthly Lag 0, 3 Lag 3 MODIS derivative [31]
LST_day Daytime land surface temperature Monthly MODIS derivative [30]
LST_delta Diurnal difference in land surface Monthly Lag 0, 1, 2, 3 Lag 0, 1, 2, 3 MODIS derivative [30]
temperature
LST_night Night-time land surface temperature Monthly MODIS derivative [30]
TCB Tasselled cap brightness; measure of land Monthly Lag 0, 2 Lag 2 MODIS derivative [32]
reflectance
TCW Tasselled cap wetness Monthly Lag 3 Lag 3 MODIS derivative [32]
TSI Temperature suitability index Monthly Lag 0, 1, 2, 3 MAP [26]
CHIRPS Climate Hazards Group Infrared Precipitation Monthly Lag 0, 1, 2, 3 Lag 0, 1, 3 CHIRPS [25]
with Station Data
CHIRPS Climate Hazards Group Infrared Precipitation with Station Data, GIS geographic information system, MAP Malaria Atlas Project, MODIS Moderate Resolution
Imaging Spectroradiometer, NOAA National Oceanic and Atmospheric Administration, SRTM Shuttle Radar Topography Mission, WWF World Wildlife Fund
Bayesian spatio-temporal model logitðplt Þ ¼ X Tlt β þ f ðYearÞ þ ϕ lt :
Parasite prevalence data (PfPR6–59mo) were modelled via a
Bayesian binomial logistic regression model with spatio- The matrix X includes an intercept and a list of environ-
temporal random effects accounting for a spatial latent mental and socio-demographic covariates known to affect
process varying with time. An integrated nested Laplace prevalence. β is the vector of regression coefficient, f(Year)
approximation [33] was adopted for model inference and is a temporal random effect accounting for differences in
prediction. The spatio-temporal random effects were years in the prevalence data and ϕlt is the spatio-
modelled using stochastic partial differential equations temporally structured random effects. The temporal ran-
[34], which represent a Matérn spatio-temporal Gaussian dom effect, f(Year), is assigned a first-order autoregressive
field as a Gaussian Markov random field via triangulation. prior distribution (AR1). Monthly variations in the data
Let Ylt, nlt and plt be the number of infected individ- are captured via the spatio-temporal process ϕlt, which
uals, the number of individuals screened and the preva- changes in time with an AR1 process modelled at two
lence of infection by P. falciparum at geocoded location semi-annual scales, as follows:
l (l = 1, …, N) for time t (t = 1, …, T). Ylt is assumed to ϕ l;t ¼ εl1 ; if t ¼ 1;
follow a binomial distribution: ϕ l;t ¼ aϕ l;t−1 þ εl;t ; if t ¼ 2;
Y lt ∼Binðplt ; nlt Þ: where a (|a| < 1) is a temporal autoregressive coefficient
and ϕl, t is the vector of the spatio-temporally structured
The prevalence of infection, plt, is modelled via a effect, which follows a multivariate normal distribution
linear regression on the logit scale: with zero mean and a spatio-temporal covariance function
Kang et al. BMC Medicine (2018) 16:71 Page 5 of 15 of the Matérn family. See [34] for more details of the spe- A range of model validation analyses were used to as- cification of the spatio-temporal random field. Since the sess the model’s goodness of fit and predictive accuracy, prevalence data were only available for certain months including the Pearson correlation of observed and pre- each year, the AR1 process (time effect) of ϕl, t is modelled dicted data, and validation runs with incremental hold- at two half-year intervals to ensure continuity: December out validation data subsets. to May and June to November. As part of the variable selection procedure, collinearity Results among the environmental covariates was examined by The model framework calculating variance inflation factors (VIFs) prior to The Bayesian spatio-temporal model was developed in fitting the statistical model to the prevalence data. A an integrated nested Laplace approximation framework stepwise selection of covariates using VIFs was under- to map the prevalence of P. falciparum malaria infection taken to make sure all VIF values are below a desired among children aged 6 to 59 months in age (PfPR6–59mo) threshold (VIF < 10 in this case). Put simply, using the across Madagascar for 2011, 2013 and 2016. Validation full set of covariates, a VIF for each variable was calcu- of the final model gave Pearson correlation coefficients lated. The variable with the single highest value was of 0.96, 0.94 and 0.83 for each prediction year, respect- removed and all VIF values with the new set of variables ively, between month-specific predicted and observed were recalculated. Then, the variable with the next high- prevalence (Additional file 2: Figure S2), justifying a est value was removed, and so on, until all VIF values were good fit of the Bayesian spatio-temporal model to the below the threshold of 10. This resulted in a set of 26 co- observed prevalence data. To assess the predictive power variates being considered for model fitting (Table 1, of the specified model, prevalence data were split into Round 1 selection). Subsequently, variable selection was two sets: a training set (Dt) and a validation set (Dv) performed using bidirectional elimination by running [37]. After being trained on Dt, the model accuracy was stepwise regressions on all model combinations using the determined by its ability to predict Dv as well as the 26 chosen covariates. The best-fitting (final) model had entire prevalence dataset. A percentage of Dv was chosen the smallest deviance information criterion value and in- to be a random sample of 10%, 20%, 30% and 40% of the cluded 18 covariates (Table 1, Round 2 selection). overall dataset. Each validation was repeated 100 times Using the final model, the prevalence of infection by P. and the model accuracy was averaged across the 100 falciparum (PfPR6–59mo) was predicted over a 30 arcse- fitted models. The box plots of cross-validated correla- cond spatial grid of 1,602,342 pixels (approximately tions appeared satisfactory (Additional file 3: Figure 1 km × 1 km spatial resolution) for each month of 2011, S3a). The cross-validated R2 had a mean of 0.78 and a 2013 and 2016. The final covariates were available at range of (0.70, 0.83) for 10% Dv; a mean of 0.72 and a monthly intervals and this allowed us to make monthly range of (0.62, 0.79) for 20% Dv; a mean of 0.65 and a predictions of prevalence. The resulting predictions range of (0.55, 0.74) for 30% Dv; and a mean of 0.58 and included the monthly mean prevalence and the monthly a range of (0.45, 0.68) for 40% Dv (Additional file 3: interquartile range (IQR) of prevalence as a measure of Figure S3b). The model was deemed accurate in its the associated prediction uncertainty. The annual mean ability to predict a wide range of validation sets prevalence (PfPR6–59mo) and mean IQR were obtained (Additional file 4: Figure S4). by averaging across the 12 monthly predictions. This Bayesian model-based geostatistical approach al- Covariates lows the smoothing of extreme rates due to small sample The combination of covariates that together best explained sizes by borrowing strength across neighbouring pixels the spatio-temporal variation in the PfPR datapoints were to improve local estimates. Essentially, the approach as- selected for inclusion in the spatio-temporal model and are sumes a positive spatial correlation between observations, listed in Tables 1 (Round 2 selection) and 2. Out of the 18 borrows more information from neighbouring pixels than predictors which informed the statistical model, eight pre- from pixels far away (in both space and time), and dictors were significant based on their 95% Bayesian cred- smooths local rates toward local, neighbouring values. ible intervals, namely LST_delta_2, TWI, PET, CHIRPS_1, The resulting smoothed prevalence estimates are robust Accessibility, Elevation, CHIRPS_0 and TCW_3. Of these and reliable while circumventing the issue of sparse data eight predictors, LST_delta_2 (diurnal variation in land [35]. The inherently hierarchical structure of the current surface temperature 2 months prior to the prediction modelling framework permits the model-based estimation month) had a positive effect on the odds of risk with odds of covariate effects, the prediction of missing data and the ratio OR = 1.1987. The odds of P. falciparum infection estimation of spatio-temporal covariance structures [36]. were negatively associated with TCW_3 (an index of sur- Model predictions are presented at the pixel level and also face wetness 3 months prior to the prediction month) with summarised by ecozone [4]. OR = 0.0025. The remaining six covariates had OR slightly
Kang et al. BMC Medicine (2018) 16:71 Page 6 of 15 Table 2 Regression coefficients and odds ratios of the predictors selected by the final model and their associated 95% Bayesian credible intervals (CI) Covariate Regression coefficient 95% CI of regression coefficient Odds ratio 95% CI of odds ratio EVI_3 1.5872 (− 0.3861, 3.5926) 4.8913 (0.6799, 36.3362) LST_delta_2 0.1812 (0.1098, 0.2527)* 1.1987 (1.1161, 1.2876)* TWI 0.0535 (0.0048, 0.1017)* 1.0550 (1.0048, 1.1071)* PET 0.0028 (0.0009, 0.0048)* 1.0028 (1.0009, 1.0048)* CHIRPS_1 0.0023 (0.0010, 0.0037)* 1.0023 (1.0010, 1.0037)* Accessibility 0.0020 (0.0009, 0.0031)* 1.0020 (1.0009, 1.0031)* CHIRPS_3 0.0004 (−0.0010, 0.0018) 1.0004 (0.9990, 1.0018) AI 0.0001 (0.0000, 0.0002) 1.0001 (1.0000, 1.0002) DistToWater 0.0000 (0.0000, 0.0001) 1.0000 (1.0000, 1.0001) Slope 0.0000 (0.0000, 0.0000) 1.0000 (1.0000, 1.0000) Elevation −0.0015 (−0.0022, − 0.0009)* 0.9985 (0.9978, 0.9991)* CHIRPS_0 −0.0020 (−0.0036, − 0.0004)* 0.9980 (0.9964, 0.9996)* LST_delta_1 −0.0111 (−0.0944, 0.0721) 0.9889 (0.9099, 1.0748) Stable_Lights_2010 −0.0274 (−0.0801, 0.0218) 0.9730 (0.9231, 1.0221) LST_delta_3 −0.0532 (−0.1173, 0.0106) 0.9482 (0.8894, 1.0106) LST_delta_0 −0.0540 (−0.1360, 0.0281) 0.9475 (0.8729, 1.0285) TCB_2 −1.7339 (−5.2983, 1.8037) 0.1765 (0.0050, 6.0672) TCW_3 −5.9884 (−10.7712, − 1.2680)* 0.0025 (0.0000, 0.2813)* Intercept −11.5150 (−15.6992, −7.5190) 0.0000 (0.0000, 0.0005) CI credible interval *Significant based on 95% Bayesian credible interval above or below 1, indicating that they did not have a large low between May and September. In contrast, the tas- influence on the odds of risk. Note that the final model also selled cap wetness (a remotely sensed transformed index included several predictors with OR being roughly one, of land surface water availability) remained relatively con- namely aridity index, distance to water and slope, which stant over time, with no significant monthly variation suggests that they did not affect the odds of P. falciparum when summarised to the national level. At the higher- infection. However, the model selection procedure that resolution ecozone level, ecological predictors show more optimised the deviance information criterion value justified extreme temporal variation, emphasising the country’s the need to retain these predictors in the model. environmental diversity. Put together, these constellations Monthly variation in the key predictors is summarised of monthly dynamic explanatory variables combined with at the national level in Fig. 2 and by the previously the static predictors explained the spatio-temporal vari- defined ecozones [4] in Additional file 5: Figure S5. ation in the input data and informed extrapolation to loca- These box plots include the temporally dynamic envir- tions and time points that were not represented in the onmental predictors that informed the statistical model MIS input datasets. over different time lags. For example, EVI with a 3- month lag indicates that the malaria prevalence predic- Spatio-temporal trends in malaria endemicity tion for January 2011 was informed by the EVI level in This paper’s primary objective was to assess changes in October 2010. LST_delta, or the diurnal variation in the prevalence of P. falciparum malaria infection among land surface temperature, was influential across the those under 5 years old (PfPR6–59mo) in Madagascar be- whole time window up to 3 months prior to the predic- tween the three surveyed years of 2011, 2013 and 2016. tion months. Predictions for January 2011, for example, The published MIS reports gave national prevalence esti- were influenced by the difference in daytime and night- mates for the 6–59 month age group of 6.2% [9], 9.1% time temperatures (LST_delta) in October, November [10] and 7.0% [11], respectively. The geostatistical model and December 2010, as well as January 2011. Seasonal developed here generated 36 monthly prediction maps of variation in the different dynamic predictors differed in endemicity (Additional file 5: Figure S5 and Additional file amplitude, with the most extreme being the CHIRPS 6: Figure S6), which are summarised into three annual dataset (a precipitation indicator), which dropped to very mean predictions (Fig. 3a–c), with ecozone-level box plot
Kang et al. BMC Medicine (2018) 16:71 Page 7 of 15 Fig. 2 National-level mean monthly PfPR6–59mo predictions, plotted alongside temporally variable predictor values. The box plot rectangles indicate the first to third quartiles (interquartile range), with the median shown as the dark line inside the box. Vertical lines correspond to the minimum and maximum values. Specified lags indicate the time points that were selected by the model as explanatory variables of PfPR6–59mo. A time lag of 0 indicates that the covariate values in the concurrent month were predictive of PfPR6–59mo, while a time lag of 3 indicates that the covariate value 3 months prior to the PfPR6–59mo prediction was predictive of PfPR6–59mo summaries of trends across the country’s eight ecozones the 3 years evaluated (30% mean annual prevalence, monthly or annually. Model uncertainty is mapped expanding in extent from 2011 into subsequent years. alongside the mean PfPR6–59mo prediction surfaces as High prevalence in this ecozone is not homogeneous the IQR of the model posterior distribution (Fig. 3d–f however, being interspersed with pockets of lower en- and Additional file 6: Figure S6b, d, f ), quantifying demicity ( 10% and with areas >30% mean annual prediction are plotted in Additional file 7: Figure S7 to PfPR6–59mo prevalence. The northern and southern tips reflect the relative variability of the predictions. have lower prevalence, typically under 5% endemicity. Together, these uncertainty maps show that while the Confidence in the model predictions was variable across absolute uncertainty (represented by the IQR) is lower the country and between years. The low-prevalence in the central highlands where prevalence is lowest, data-rich highlands were modelled with greater confi- when adjusted to the mean prediction values, confi- dence than the more heterogeneous coastal regions, dence in the prediction is strongest in coastal areas where survey data indicated spatially variable rates of in- where prevalence is higher. fection (Figs. 1 and 3d–f ). The highest model uncer- Malaria endemicity remains spatially heterogeneous tainty in all 3 years was in the north-eastern region of across Madagascar (Figs. 3 and 4), with the central high- Sava, an area previously characterised as a malaria para- lands consistently having the lowest endemicity across site transmission hotspot [38]. As seen across the whole
Kang et al. BMC Medicine (2018) 16:71 Page 8 of 15 Fig. 3 Predicted annual mean PfPR among children 6 to 59 months in age for 2011 (a), 2013 (b) and 2016 (c). d–f The corresponding map uncertainty (quantified as the prediction interquartile range). Values are mapped at 1 × 1 km pixel resolution. g–i National population breakdown by endemicity class, using population values based on WorldPop’s Whole Continent UN-adjusted Population Count datasets for Africa for 2010, 2015 and 2020. Estimates for 2011, 2013 and 2016 were created by linear interpolation of the bookending quinquennial rasters country after 2011, the higher observed prevalence was with a standard deviation of 34.0% around the 127% mean associated with a larger IQR, reflecting the wider range increase. All ecozones experienced at least a doubling in of potential prevalence values. prevalence of PfPR6–59mo between 2011 and 2016, with At the national level, the three mean annual maps were the smallest proportional change being in the south- all significantly different from one another (Wilcoxon rank east (100.4%; Fig. 4 and Additional file 5: Figure S5d) sum test, p < 2.2 × 10-16). PfPR6–59mo more than doubled and the highest up to 157.4% in the central highlands, between 2011 and 2016 (127% increase across the mean where prevalence nevertheless remained the lowest annual maps; Fig. 5a and c), despite a 23% decrease within (Fig. 4 and Additional file 5: Figure S5h). The south that window between 2013 and 2016 (Fig. 5b and d). ecozone also saw an important increase of 142.5% over Changes in endemicity were spatially variable across the the 6-year period (Fig. 4 and Additional file 5: Figure period examined. The magnitude of change across the S5g). The 23% mean fall in prevalence from 2013 to country between 2011 and 2016 was highly heterogeneous, 2016 was fairly consistent across the whole country
Kang et al. BMC Medicine (2018) 16:71 Page 9 of 15
Fig. 4 Box plots of predicted monthly PfPR6–59mo by ecozone for 2011, 2013 and 2016. Ecozone extents are shown in Fig. 1 [4]
(standard deviation 7.4%), and at the ecozone level, lowest endemicity areas, with almost half (42%) of the
ranged from a 20.7% decrease in the central highlands population in very low endemicity areas (10% in 2011. The highest endemicity
of 10% to 50% between 2013 and 2016. In contrast, 27% year, 2013, was predicted to have 32.5% of the popula-
of pixels increased between 0% and 100%, and 71% tion in areas where endemicity was >10%, a threefold
more than doubled in prevalence over the full study increase from 2011. By 2016, this had dropped to 25.3%
period from 2011 to 2016. of the population being in areas of >10% prevalence,
with a corresponding increase in the proportion living
Trends in population exposure in very low (Kang et al. BMC Medicine (2018) 16:71 Page 10 of 15
Fig. 5 Percentage changes in predicted PfPR among children 6 to 59 months old across the three MIS time points: a from 2011 to 2016 and b from
2013 to 2016. Histograms of pixel-level change c from 2011 to 2016 and d from 2013 to 2016. Positive % change indicates an increase in prevalence,
while negative % change is a decrease
low endemicity areas where most of the population lives, important increasing trends since 2011 despite a relatively
and notable increases in the population exposed to the small fall in prevalence between 2013 and 2016.
highest end of the endemicity spectrum. A decrease of ex-
posure levels from 2013 to 2016 was evident, but much Comparison with reported MIS results
less pronounced than the increase between 2011 and The results presented in MIS reports are summaries of
2013, meaning that the exposure levels in 2016 were con- the raw survey data, with individuals weighted in such a
siderably worse than in 2011. way as to ensure appropriate national representation in
the overall summary statistics. However, the results are
Discussion not adjusted for the temporal lag in survey dates across
Three MIS studies have taken place in Madagascar since years (Additional file 1: Figure S1). By 2016, the survey
2011, providing a valuable source of information about the time window was pushed to 2 months later than in
status of malaria from a standardised data collection proto- 2011, corresponding to a delay away from the peak
col applied across a nationally representative set of loca- transmission period in most regions (Additional file 3:
tions. A broad range of indicator metrics are included in Figure S3 and Additional file 4: Figure S4). The model
the MIS reports. This present study looks at the parasit- presented here specifically accounts for this temporal
ology data in particular to assess how the prevalence of lag, allowing for meaningful comparisons between years.
malaria infection in children 6 to 59 months old has chan- Seasonal trends in transmission [4] mean that a de-
ged in recent years. Given the limitations of the routine layed survey period will underestimate endemicity
health data reporting chain discussed previously [4], this relative to the original survey time period. This is
study provides a complementary perspective into the recent reflected in the raw MIS results, which suggest a 6.2%
status of malaria endemicity in Madagascar, identifying national PfPR6–59mo in 2011 and 7% in 2016, a 13%Kang et al. BMC Medicine (2018) 16:71 Page 11 of 15 increase over that period. In contrast, the spatio- burden of disease and an important increase in burden par- temporal model presented here identifies a 127% in- ticularly along the west coast. A major increase in clinical crease, resulting in an estimated national mean 9.3% cases was reported between 2014 and 2015, which subse- annual PfPR6–59mo in 2016. quently reduced in 2016 (the third MIS year) following a The MIS specifically excludes three highland districts mass distribution of bed nets treated with insecticide at the considered to be malaria-free (the cities of Antananarivo- end of 2015 [41]. The longitudinal nature of the data from Renivohitra, Antsirabe I and Fianarantsoa I) as well as the health metrics information system allows extreme communes above 1500 m in altitude. Despite not being events to be identified, such as epidemics, which may dras- represented in the mapping input dataset, model predic- tically affect the annual case totals, as in 2015, but which tions are derived for these areas, which then inform the may not be distinguishable as exceptional, or even captured ecozone- and national-level summaries. Without prior by cross-sectional surveys that assess prevalence at isolated exclusion of these zones, the model-based approach risked time points. over-inflating the estimates of endemicity. Environmental A 2017 World Health Organization (WHO) report covariates, however, appear to have discerned the unsuit- estimates that only around 31% of all clinical cases are ability of the highland urban habitat, and the malaria-free reported through the surveillance system to the central districts are predicted to have
Kang et al. BMC Medicine (2018) 16:71 Page 12 of 15 NMCP to target control measures optimally. Elimination used allowed robust predictions of malaria prevalence, requires both the reduction of the parasite reservoir and so this was not considered a major limitation to the the prevention of transmission. Prevalence data alone mapping model presented here. cannot fully characterise the local epidemiology without Madagascar is noted for its mosaic landscape of eco- a parallel understanding of a site’s historical context, the logical habitats, with land cover varying across short dis- significance of imported cases, the impact of recent tances [44]. The critical importance of the environmental control efforts, entomological and host behavioural/gen- covariates in the modelling process is evident from the etic factors, and so on [43]. A practical interpretation of methods described here, with a strong predictive role asso- the prevalence map, therefore, requires insight into the ciated with vegetation cover that explains differences in underlying factors driving the observed infection rates. malaria prevalence between different areas (Table 2). The The map suite presented here is one component in covariates associated with each MIS site describe the local understanding malaria in Madagascar, but ought to be conditions associated with the observed malaria prevalence interpreted in association with a broader set of malario- at the time and location of sampling. However, MIS data- metric data when used to determine control intervention sets are geopositioned with a deliberate degree of spatial policy. Malaria transmission is also highly dynamic both uncertainty (displacement) to promote anonymity of up to spatially and temporally, meaning that predicted maps 2 km in urban areas, 5 km in rural areas and 10 km for 1% based on single time-point snapshots of prevalence may of rural points [45]. This spatial displacement, therefore, simplify the true underlying situation. Maps such as those introduces uncertainty into the associations between re- presented here provide insight into general trends, and ported PfPR and their attributed covariate values, which the validation statistics presented in Additional file 2: could impact the model’s predictions. For this study, we as- Figure S2, Additional file 3: Figure S3 and Additional file 4: sumed that while the island is ecologically heterogeneous, Figure S4 indicate the level of variability that might be ex- the impact of this spatial uncertainty will be acceptably low pected around these predictions. (with ecological conditions similar for most points even at Malaria prevalence is strongly influenced by interven- a distance of 5 km or 10 km), and similar enough to allow tion coverage levels, including rates of insecticide- a prevalence signal to be identified. The model validation treated net (ITN) ownership [39] and treatment seeking statistics corroborate this assumption. [13, 14]. Including these covariates in the modelling A further limitation of the MIS datasets that informed framework could help inform the model about the pat- the current mapping analysis stems from their sampling terns of endemicity but this was not considered feasible design and sample sizes. The broad range of indicators in this present analysis. The coverage of indoor residual included in the MIS activities present conflicting de- spraying and ITN use rates, for instance, have complex mands on sample sizes, which are further constrained by non-linear relationships with malaria prevalence. For financial and logistical considerations. Sample sizes can- example, indoor residual spraying in Madagascar is care- not, therefore, be optimised for all indicators, but in- fully targeted to the highest (as an emergency response stead are focussed on a limited number of these. A to reduce mortality during outbreaks) and lowest (pre- recent retrospective model-based analysis of the 2011 vention of reintroduction and subsequent autochthon- and 2013 Madagascar MIS datasets estimated that sam- ous transmission) endemicity districts only [8]. ITN ple sizes were under-powered by 17% and 36%, respect- coverage is strongly skewed to areas where malaria is en- ively, to reach effective sizes for infection prevalence demic. The highland areas into which malaria is mainly rates [46]. This was based on rapid diagnostic test results imported are not covered by routine ITN distribution. and not microscopy (as considered in this present study), The limited temporal window considered in this present meaning that it may be a slight overestimate. The ap- study does not allow for the protective effect of high proach followed here, namely to consider samples from ITN coverage to be learnt by the model, and instead the the three MIS datasets within a common Bayesian hier- coarse learnt association is that ITN coverage increases archical modelling framework and to draw on a broad as prevalence increases. Treatment seeking was excluded range of associated covariate surfaces, provides a solution for reasons of sample sizes. MIS data on treatment seek- to sample size limitations. ing from individual cluster locations are limited to Furthermore, at very low transmission levels, such as mothers with infants who suffered from fever in the 2 in the central highlands and parts of the central fringe weeks preceding the MIS interviews. Sample sizes at the regions of Madagascar, microscopy-based cross-sectional cluster level are, therefore, very small, producing spuri- surveys of parasite prevalence risk are underpowered to ous results when analysed at the high resolution of the detect rare and low parasitaemia infections adequately present analysis. Despite these barriers to including [47]. In these areas, higher sensitivity diagnostics (such intervention covariates in the model, the suite of envir- as nucleic-acid amplification-based approaches) and onmental and socio-demographic variables that were alternative indicators (such as serological markers) are
Kang et al. BMC Medicine (2018) 16:71 Page 13 of 15
required to monitor malaria epidemiological trends geographically progressive elimination. Prevalence maps,
more effectively [47–50]. Particular interest is in the use as presented here, represent one component of monitor-
of serological panels targeting both short- and long- ing progress towards those goals.
lasting antigenic responses [51]. When applied to appro-
priate sentinel populations [52, 53], these tools can be Additional files
powerful probes of changing transmission intensity or
reintroduction events in the pre-elimination setting Additional file 1: Figure S1. Sample screening during the three MIS
where the majority of infections at the time of survey events in Madagascar, showing the progressive delay in the sampling
may be microscopically sub-patent [54]. time window. a Overall bar plots of sampling months. b–d Maps by
sampling month by cluster location. (PNG 1353 kb)
Additional file 2: Figure S2. Month-specific correlations between observed
raw MIS prevalence values and the model predictions for each annual map.
Conclusions Pearson correlation coefficients are shown on each plot. (PNG 32 kb)
Malaria remains an important health problem in Additional file 3: Figure S3. Box plots of a cross-validated Pearson
Madagascar, with the prevalence of infection more than correlation coefficients and b cross-validated R2, based on 100 randomly
sampled validation sets. (PNG 26 kb)
doubling between 2011 and 2016 (a mean increase of
Additional file 4: Figure S4. Observed prevalence against predicted
127%). Across the three survey time points, 2011 to 2013 prevalence averaged across 100 randomly sampled validation sets.
saw the greatest PfPR6–59mo increase, followed by a reduc- (PNG 63 kb)
tion to 2016 (mean reduction of 23%). However, while the Additional file 5: Figure S5. Summary box plots of predicted monthly
whole population is at risk of infection, prevalence was PfPR6–59mo by ecozone, plotted alongside temporally variable predictor
values. The box plot rectangles indicate the first to third quartiles
lower in higher density areas, with 26.7% of the population (interquartile range), with the median shown as the dark line inside the
in 2016 living in pre-elimination areas, where prevalence box. Vertical lines correspond to the minimum and maximum values.
wasKang et al. BMC Medicine (2018) 16:71 Page 14 of 15
Authors’ contributions 12. Barmania S. Madagascar's health challenges. Lancet. 2015;386:729–30.
REH, SYK, KEB, and PWG conceived the study. KEB and HG prepared the 13. Cibulskis RE, Aregawi M, Williams R, Otten M, Dye C. Worldwide incidence
input data and covariate surfaces. SYK developed the statistical model. DW, of malaria in 2009: estimates, time trends, and a critique of methods. PLoS
EC and PWG advised on the analysis. AR and MR were involved with the Med. 2011;8:e1001142.
original MIS activities and generation of parasite prevalence data. REH wrote 14. Battle KE, Bisanzio D, Gibson HS, Bhatt S, Cameron E, Weiss DJ, et al.
the first draft of the manuscript with SYK and KEB. All authors contributed to Treatment-seeking rates in malaria endemic countries. Malar J. 2016;15:20.
the interpretation of results. All authors read and approved the final 15. Weiss DJ, Mappin B, Dalrymple U, Bhatt S, Cameron E, Hay SI, et al. Re-
manuscript. examining environmental correlates of Plasmodium falciparum malaria
endemicity: a data-intensive variable selection approach. Malar J. 2015;14:68.
Ethics approval and consent to participate 16. National Oceanic and Atmospheric Administration (NOAA) National Centers
Not applicable. for Environmental Information. Defense Meteorological Satellite Program
(DMSP): Data Archive, Research, and Products. http://ngdc.noaa.gov/eog/
Competing interests dmsp.html. Accessed Sept 2017.
The authors declare that they have no competing interests. 17. Noor AM, Alegana VA, Gething PW, Tatem AJ, Snow RW. Using remotely
sensed night-time light as a proxy for poverty in Africa. Popul Health
Metrics. 2008;6:5.
Publisher’s Note 18. Nelson A. Travel time to major cities: A global map of Accessibility. Ispra:
Springer Nature remains neutral with regard to jurisdictional claims in
Global Environmental Monitoring Unit - Joint Research Centre of the
published maps and institutional affiliations.
European Commission; 2008. http://forobs.jrc.ec.europa.eu/products/gam/.
Accessed Sept 2017.
Author details
1 19. Farr TG, Rosen PA, Caro E, Crippen R, Duren R, Hensley S, et al. The Shuttle
Malaria Atlas Project, Oxford Big Data Institute, Nuffield Department of
Radar Topography Mission. Rev Geophys. 2007;45:RG2004.
Medicine, University of Oxford, Oxford, UK. 2National Malaria Control
20. ESRI. ArcGIS Desktop 10.5.1. Redlands: Environmental Systems Resource
Programme, Ministry of Health, Antananarivo, Madagascar. 3University of
Institute; 2017.
Antananarivo, Antananarivo, Madagascar. 4Institut Pasteur de Madagascar,
21. HydroSHEDS database. http://hydrosheds.org/. Accessed Sept 2017.
Antananarivo, Madagascar. 5Faculté des Sciences, Université de Toliara,
Toliara, Madagascar. 6Center for Global Health and Diseases, Case Western 22. World Wildlife Fund (WWF). Global Lakes and Wetlands Database.
Reserve University, Cleveland, OH, USA. Washington, DC; 2004. https://www.worldwildlife.org/pages/global-lakes-
and-wetlands-database. Accessed Sept 2017.
Received: 12 February 2018 Accepted: 24 April 2018 23. Trabucco A, Zomer RJ. Global aridity index (global-aridity) and global
potential evapo-transpiration (global-PET) geospatial database. CGIAR
Consortium for Spatial Information ed. 2009. Available online from the
References CGIAR-CSI GeoPortal at: http://www.cgiar-csi.org/. Accessed Sept 2017.
1. Joncour G. The fight against malaria in Madagascar. Bull World Health 24. Linard C, Gilbert M, Snow RW, Noor AM, Tatem AJ. Population
Organ. 1956;15:711–23. distribution, settlement patterns and accessibility across Africa in 2010.
2. Mouchet J, Blanchy S. Particularités et stratification du paludisme à PLoS One. 2012;7:e31743.
Madagascar. Santé. 1995;5:386–8. 25. Funk CC, Peterson PJ, Landsfeld MF, Pedreros DH, Verdin JP, Rowland
3. Randrianarivelojosia M, Raveloson A, Randriamanantena A, Juliano JJ, JD, et al. A quasi-global precipitation time series for drought
Andrianjafy T, Raharimalala LA, et al. Lessons learnt from the six decades of monitoring: U.S. Geological Survey Data Series, vol. 832; 2014. https://
chloroquine use (1945-2005) to control malaria in Madagascar. Trans R Soc doi.org/10.3133/ds832.
Trop Med Hyg. 2009;103:3–10. 26. Weiss DJ, Bhatt S, Mappin B, Van Boeckel TP, Smith DL, Hay SI, et al.
4. Howes RE, Mioramalala SA, Ramiranirina B, Franchard T, Rakotorahalahy AJ, Air temperature suitability for Plasmodium falciparum malaria
Bisanzio D, et al. Contemporary epidemiological overview of malaria in transmission in Africa 2000-2012: a high-resolution spatiotemporal
Madagascar: operational utility of reported routine case data for malaria prediction. Malar J. 2014;13:171.
control planning. Malar J. 2016;15:502. 27. Gething PW, Van Boeckel TP, Smith DL, Guerra CA, Patil AP, Snow RW, et al.
5. Albonico M, De Giorgi F, Razanakolona J, Raveloson A, Sabatinelli G, Pietra V, Modelling the global constraints of temperature on transmission of
et al. Control of epidemic malaria on the highlands of Madagascar. Plasmodium falciparum and P. vivax. Parasit Vectors. 2011;4:92.
Parassitologia. 1999;41:373–6. 28. NASA. Moderate Resolution Imaging Spectroradiometer (MODIS). https://
6. Laventure S, Mouchet J, Blanchy S, Marrama L, Rabarison P, modis.gsfc.nasa.gov/. Accessed Sept 2017.
Andrianaivolambo L, et al. Rice: source of life and death on the plateaux of 29. Weiss DJ, Atkinson PM, Bhatt S, Mappin B, Hay SI, Gething PW. An effective
Madagascar. Sante. 1996;6:79–86. approach for gap-filling continental scale remotely sensed time-series. ISPRS
7. Girond F, Randrianasolo L, Randriamampionona L, Rakotomanana F, J Photogramm Remote Sens. 2014;98:106–18.
Randrianarivelojosia M, Ratsitorahina M, et al. Analysing trends and 30. NASA Earth Observations (NEO). Average land surface temperature. http://
forecasting malaria epidemics in Madagascar using a sentinel surveillance neo.sci.gsfc.nasa.gov/view.php?datasetId=MOD_LSTD_CLIM_M. Accessed
network: a web-based application. Malar J. 2017;16:72. Sept 2017.
8. National Malaria Control Programme of Madagascar. National strategic plan 31. NASA Earth Data. MODIS (MOD 13) - Gridded Vegetation Indices (NDVI and
for malaria control in Madagascar 2018–2022. Progressive malaria EVI). http://modis.gsfc.nasa.gov/data/dataprod/dataproducts.php?MOD_
elimination from Madagascar. (2017). NUMBER=13. Accessed Sept 2017.
9. Institut National de la Statistique (INSTAT), Programme National de Lutte 32. NASA Earth Data. Land Processes Distributed Active Archive Center. https://
contre le Paludisme (PNLP), and ICF International. Madagascar Malaria lpdaac.usgs.gov/dataset_discovery/modis/modis_products_table/mcd43b5.
Indicator Survey 2011 [Enquête sur les Indicateurs du Paludisme (EIPM)]. Accessed Sept 2017.
Calverton: INSTAT, PNLP, and ICF International; 2011. 33. Rue H, Martino S, Chopin N. Approximate Bayesian inference for latent
10. Institut National de la Statistique (INSTAT), Programme National de Lutte Gaussian models by using integrated nested Laplace approximations. J
contre le Paludisme (PNLP), Institut Pasteur de Madagascar (IPM), and ICF Royal Stat Soc Series B Stat Methodo. 2009;71:319–92.
International. Madagascar Malaria Indicator Survey 2013 [Enquête sur les 34. Cameletti M, Lindgren F, Simpson D, Rue H. Spatio-temporal modeling of
Indicateurs du Paludisme (EIPM)]. Calverton: INSTAT, PNLP, IPM and ICF particulate matter concentration through the SPDE approach. Adv Stat Ana.
International; 2013. 2013;97:109–31.
11. Institut National de la Statistique (INSTAT), Programme National de Lutte 35. Clayton D, Kaldor J. Empirical Bayes Estimates of Age-Standardized Relative
contre le Paludisme (PNLP), Institut Pasteur de Madagascar (IPM), and ICF Risks for Use in Disease Mapping. Biometrics. 1987;43:671–81.
International. Madagascar Malaria Indicator Survey 2016 [Enquête sur les 36. Waller L, Carlin B. Disease mapping. In: Gelfand AE, Diggle P, Guttorp P,
Indicateurs du Paludisme (EIPM)]. Calverton: INSTAT, PNLP, IPM and ICF Fuentes M, editors. Handbook of Spatial Statistics. Boca-Raton: CRC Press;
International; 2016. 2010.Kang et al. BMC Medicine (2018) 16:71 Page 15 of 15
37. Gething PW, Patil AP, Smith DL, Guerra CA, Elyazar IR, Johnston GL, et al. A
new world malaria map: Plasmodium falciparum endemicity in 2010. Malar J.
2011;10:378.
38. Rice BL, Golden CD, Anjaranirina EJ, Botelho CM, Volkman SK, Hartl DL.
Genetic evidence that the Makira region in northeastern Madagascar is a
hotspot of malaria transmission. Malar J. 2016;15:596.
39. Bhatt S, Weiss DJ, Cameron E, Bisanzio D, Mappin B, Dalrymple U, et al. The
effect of malaria control on Plasmodium falciparum in Africa between 2000
and 2015. Nature. 2015;526:207–11.
40. WHO. World Malaria Report 2017. Geneva: World Health Organization; 2017.
41. National Malaria Control Programme of Madagascar. Malaria Programme
Review 2017 [Revue du Programme National de Lutte contre le Paludisme].
Antananarivo: Ministry of Health of Madagascar; 2017.
42. Randriamaherijaona S, Raharinjatovo J, Boyer S. Durability monitoring of
long-lasting insecticidal (mosquito) nets (LLINs) in Madagascar: physical
integrity and insecticidal activity. Parasit Vectors. 2017;10:564.
43. Cohen JM, Le Menach A, Pothin E, Eisele TP, Gething PW, Eckhoff PA, et al.
Mapping multiple components of malaria risk for improved targeting of
elimination interventions. Malar J. 2017;16:459.
44. Virah-Sawmy M, Gillson L, Willis KJ. How does spatial heterogeneity
influence resilience to climatic changes? Ecological dynamics in southeast
Madagascar. Ecol Monogr. 2009;79:557–74.
45. Burgert CR, Colston J, Roy T, Zachary B. Geographic displacement procedure
and georeferenced data release policy for the demographic and health
surveys. In: DHS Spatial Analysis Reports No 7. Calverton: ICF International;
2013. https://dhsprogram.com/pubs/pdf/SAR7/SAR7.pdf.
46. Alegana VA, Wright J, Bosco C, Okiro EA, Atkinson PM, Snow RW, et al. Malaria
prevalence metrics in low- and middle-income countries: an assessment of
precision in nationally-representative surveys. Malar J. 2017;16:475.
47. Dalrymple U, Mappin B, Gething PW. Malaria mapping: understanding the
global endemicity of falciparum and vivax malaria. BMC Med. 2015;13:140.
48. Yalew WG, Pal S, Bansil P, Dabbs R, Tetteh K, Guinovart C, et al. Current and
cumulative malaria infections in a setting embarking on elimination:
Amhara, Ethiopia. Malar J. 2017;16:242.
49. Corran P, Coleman P, Riley E, Drakeley C. Serology: a robust indicator of
malaria transmission intensity? Trends Parasitol. 2007;23:575–82.
50. Zimmerman PA, Howes RE. Malaria diagnosis for malaria elimination. Curr
Opin Infect Dis. 2015;28:446–54.
51. Helb DA, Tetteh KK, Felgner PL, Skinner J, Hubbard A, Arinaitwe E, et al.
Novel serologic biomarkers provide accurate estimates of recent
Plasmodium falciparum exposure for individuals and communities. Proc Natl
Acad Sci U S A. 2015;112:E4438–47.
52. Ray K, Upreti HB, Yadav RN, Sharma MC, Mukherjee AK, Sharma GK.
Evaluation of serology as a tool for malaria surveillance in East Champaran
District of Bihar, India. Ann Trop Med Parasitol. 1988;82:225–8.
53. Elliott SR, Fowkes FJ, Richards JS, Reiling L, Drew DR, Beeson JG. Research
priorities for the development and implementation of serological tools for
malaria surveillance. F1000Prime Rep. 2014;6:100.
54. Stresman GH, Stevenson JC, Ngwu N, Marube E, Owaga C, Drakeley C, et al.
High levels of asymptomatic and subpatent Plasmodium falciparum parasite
carriage at health facilities in an area of heterogeneous malaria transmission
intensity in the Kenyan highlands. Am J Trop Med Hyg. 2014;91:1101–8.
55. The World Bank. Madagascar: Measuring the Impact of the Political Crisis.
Washington DC; 2013. http://www.worldbank.org/en/news/feature/2013/06/
05/madagascar-measuring-the-impact-of-the-political-crisis. Accessed 31 Jan
2016.
56. Baird JK. Malaria control by commodities without practical malariology. BMC
Public Health. 2017;17:590.You can also read
